Lateralized Learning to Solve Complex Problems

Artificial intelligence systems have become proficient at linking environmental features to targets to describe simple patterns in data. However, these systems can struggle with many real-world problems that entail hierarchical patterns within patterns, for example, in recognizing object ontologies where one object is made-up of other objects. Although it is possible to capture such complex structures by utilizing state-of-the-art deep networks, the knowledge is often stored in layers that do not take advantage of the potential benefits provided by reusing patterns within a layer of the system.

Biological nervous systems can learn knowledge from simple and small-scale problems and then apply it to resolve more complex and large-scale problems in similar and related domains. However, rudimentary attempts to apply this transfer learning in artificial intelligence systems have struggled. This may be due to the homogeneous nature of their knowledge representation. The current understanding of the learning mechanisms in the brains of human and non-human animals can be used as inspiration to improve learning in artificial agents. Research into lateral asymmetry of the brain shows that it enables modular learning at different levels of abstraction that facilitate transfer between tasks.

The proposed thesis is that an artificial intelligence system that enables lateralization and modular learning at different levels of abstraction has the ability to solve complex hierarchical problems that a similar homogeneous system can not. The comprehensive goal of this thesis is to accomplish lateralized learning, inspired by the principles of biological intelligence, in artificial intelligence systems. The objectives are to show that lateralization and modular learning assist the novel systems to encapsulate the underlying knowledge patterns in the form of building blocks of knowledge. These building blocks of knowledge are to be tested on analyzable Boolean tasks as well as practical computer vision and navigation tasks. Academic contributions are related to the novel methods of the linking, transfer, and sharing of learned knowledge which are based on the analogous strategies of the brain.

This thesis proposes a general framework for lateralized artificial intelligence systems. The novel lateralized framework spans key aspects of knowledge perception, knowledge representation and utilization, and patterns of connectivity. It determines the essential functionality, critical methods, and associated parameters that are required to be incorporated into an artificial intelligence system to behave as a lateralized artificial intelligence system.

This thesis creates a novel evolutionary machine learning system, by adapting the lateralized framework, to obtain a proof-of-concept of the lateralized approach. Considering the same problem at different levels of abstraction enables the novel system to reframe a complex problem as a simple problem and efficiently resolve it. The results on analyzable Boolean tasks show that the problems that contain a natural hierarchy of patterns are solved to a scale that exceeds previous work (i.e. 18-bit hierarchical multiplexer problem), and reusing learned general patterns as constituents for future problems advances transfer learning (e.g. n-bit parity problem effectively becomes a sequence of 2-bit parity problems).

This thesis creates a novel lateralized artificial intelligence system, by adapting the lateralized framework, that shows robustness in a real-world domain that includes uncertainty, noise, and irrelevant and redundant data. The results of image classification tasks show that the lateralized system efficiently learns hierarchical distributions of knowledge, demonstrating performance that is similar to (or better than) other state-of-the-art deep systems as it reasons using multiple representations. Crucially, the novel system outperformed all the state-of-the-art deep models for the classification (binary classes) of normal and adversarial images by 0.43%-2.56% and 2.15%-25.84%, respectively. This thesis creates another novel multi-class lateralized system for computer vision problems to show that the lateralized approach can be scaled and not limited to learning classifier systems.

Both the Boolean and computer vision problems are single step problems in the spatial domain. However, most biological tasks, which exhibit heterogeneity, are temporal in nature. This thesis creates a novel frame-of-reference based artificial intelligence system, by adapting the lateralized framework, to address perceptual aliasing in multi-step decision making tasks. Considering aliased states at a constituent level enables the novel system to place them appropriately in holistic level policies. Consequently, the novel system transforms a non-Markov environment into a deterministic environment and efficiently resolves it. Experimental results show that the novel system effectively solves complex aliasing patterns in non-Markov environments that have been challenging to artificial agents. For example, the novel system utilizes only 6.5, 3.71, and 3.22 steps to resolve Maze10, Littman57, and Woods102, respectively.

A final contribution of this work is to obtain evidence of the benefits/costs of lateralization from artificial intelligence in order to inform cognitive neuroscience. Given that lateralization is ubiquitous in brains, evolutionary benefits can be assumed, at least in some domains. But that does not mean those benefits extend to all domains. The cognitive neuroscience research community has been struggling to determine the trade-off between the benefits and costs of lateralization. It has been hypothesized that lateralization has benefits that may counterbalance its costs. Lateralization has been associated with both poor and good performance. This thesis demonstrates the value of viable artificial systems for testing the costs and benefits of lateralization in biological systems.